Large Vision-Language Models (LVLMs) have excelled in joint visual and language understanding, particularly in image captioning. However, LVLMs remain prone to object hallucination, describing non-existent entities in the given image. Previous remedies, including dataset-enhanced supervised training and reinforcement learning from human feedback, show promise but rely heavily on costly human involvement, limiting scalability. To address this, we propose a reinforcement learning framework to mitigate object hallucination in LVLM-based image captioning, driven entirely by natural language processing (NLP) metrics. While optimizing reinforcement learning solely on NLP metrics offers scalability, it also introduces pitfalls: reward hacking, sparse feedback, and costly training. Our framework sidesteps these by introducing an Object Consistency Score (OCS) capturing both correctness and coverage of objects, incorporating KL-regularization for denser token-level guidance, and designing a lightweight PPO variant that fine-tunes adapters atop a frozen LVLM backbone, achieving efficiency and stability. Extensive experiments on the InstructBLIP baseline demonstrate the effectiveness and generalizability of our framework, achieving a substantial hallucination reduction of up to 41% on the COCO dataset and 29% on the Visual Genome. Furthermore, the framework can enhance caption quality by incorporating complementary evaluation metrics such as BERTScore and METEOR. These results reveal a broader insight in multimodal learning: LVLMs can effectively self-correct hallucinations using metric feedback, motivating deeper exploration into scalable reinforcement frameworks that significantly reduce human effort.
Figures & tables
Fig. 1 : Quantitative and qualitative comparison between Base (InstructBLIP) and Ours : Chart (a) shows a significant 41% reduction in object hallucination on the COCO dataset using Ours . Figure (b) presents an example where the Base model produces a caption with substantial object hallucination, while the Ours model provides an accurate description without hallucinated objects.
Fig. 2 : Detailed architecture of our framework. Specifically, the Policy Network is crafted by augmenting the shared LVLM with delicately learnable soft prompts. Meanwhile, the Value Network is formed by replacing the LLM Head with a Linear Value head. Notably, all parameters of the LVLM remain shared and frozen, with only a very small fraction (less than 0.01% LVLM weight) of trainable parameters added to the LVLM for the meticulous modeling of the policy network and value network.
Fig. 3 : Simplified overview of the framework’s network components: policy network, value network, and reference network. All networks share the same frozen LVLM as its foundation. The reference network mirrors the LVLM identically, while the value and policy networks incorporate a lightweight adapter into the shared LVLM.
Type
Prompt
InstructBLIP
mPLUG-Owl
CHAIRi (%) ↓
CHAIRs (%) ↓
CHAIRi (%) ↓
CHAIRs (%) ↓
Short
Generate a short caption of the image.
2.43
3.13
22.81
60.55
Create a textual summary for the image.
4.95
6.51
22.98
61.33
Long
Provide a detailed description of the image.
27.01
60.91
26.03
71.39
Create a detailed textual summary for the image.
25.80
59.11
24.25
66.31
TABLE I : Object Hallucination, gauged by CHAIRs and CHAIRi metrics, across diverse caption types paired with specific input prompts in the COCO test set. These prompts are designed to elicit both short and long captions. Two distinct methods are illustrated: InstructBLIP and mPLUG-Owl.
Method
CHAIRi(%)↓
CHAIRs(%)↓
Coverage(%)↑
mPLUG-Owl
26.2
70.5
76.7
LLaVA
22.5
62.7
79.6
LlaVA-Next
13.0
32.0
68.2
InstructBLIP
25.8
59.1
71.3
Teacher
7.5
36.4
NA
CoT
7.8
35.7
NA
TABLE II : Performance of object hallucination on the COCO dataset. The first row presents the non-fine-tuned LVLM baselines. Starting from the second row, all methods are built upon the InstructBLIP baseline. Specifically, the second row includes hallucination mitigation approaches reported in [ 98 ] . We additionally report Coverage (i.e., the recall percentage of ground-truth objects mentioned) to quantify recall preservation. The third row presents the results of our Hard Prompting strategy, while the final row reports the performance of the baseline fine-tuned using our RL framework.
Method
Captioning Quality
SPICE (%)↑
BLEU (%)↑
BERTScore (%)↑
mPLUG-Owl
12.5
2.7
87.4
LLaVA
13.5
3.0
87.8
LLaVA-Next
13.6
2.9
87.5
InstructBLIP
10.9
1.1
85.8
Hard Prompt
11.1
1.0
85.9
TABLE III : Captioning quality is evaluated using NLP metrics, comparing our approach to other baseline. Ours uses only OCS and KL divergence, while Ours+ incorporates additional metrics: METEOR and BERTScore in the reward function.
Model
CCEvali↓
CCEvals↓
InstructBLIP
22.3
72.0
Ours
15.6
43.0
TABLE IV : Performance on the Visual Genome dataset. Note that for Visual Genome, CCEval is selected as the performance metric instead of CHAIR. CCEval provides a more accurate evaluation by extending CHAIR’s methodology using GPT-assisted parsing for object extraction, specifically tailored to the Visual Genome dataset.
Model
CHAIRi↓
CHAIRs↓
BERTScore↑
BLEU↑
LLaVA
22.5
62.7
87.4
3.0
LLaVA+Ours
7.7
22.2
87.5
3.1
TABLE V : Comparison of LLaVA baseline and LLaVa fine-tuned using our method.
Fig. 4 : Qualitative comparison between the original InstructBLIP model ( Base ) and the fine-tuned InstructBLIP model trained with the proposed framework ( Ours ) on the COCO dataset. Hallucinated objects are highlighted in red . Compared with the baseline, our method generates captions that are more grounded in the visual content and significantly reduces object hallucinations.
Reward
Pre (%) ↑
Rec (%) ↑
CHAIRi (%) ↓
CHAIRs (%) ↓
Base
72.9
71.3
27.1
60.9
CHAIR
93.7
20.6
6.3
14.4
OCS
93.2
70.2
6.8
17.8
TABLE VI : Comparison of Precision ( Pre ) and Recall ( Rec ) between using CHAIR and OCS in the reward function.
Base
OCS
BERTScore
METEOR
CHAIRi
BERTScore
BLUE
✓
25.8
85.8
1.1
✓
✓
6.8
86.9
1.5
✓
✓
✓
6.9
90.5
1.8
✓
✓
✓
✓
6.9
92.4
6.6
TABLE VII : The ablation studies examining the impact of OCS, BERTscore and METEOR metrics on the COCO test set
Model
Detection mAP
Top1-Acc
CHAIRi
CHAIRs
Baseline
-
-
25.8
59.1
YOLOV8
53.9
79.0
7.8
25.5
YOLOV11
54.7
79.5
7.8
22.3
GT
-
-
6.8
17.8
TABLE VIII : Effect of object detection accuracy on hallucination reduction. YOLO-based detections greatly reduce hallucination compared to the baseline, with YOLOv11x performing slightly better; ground-truth objects remain most effective
Figure 13Figure 14Figure 15Figure 16Figure 17
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Parameters
Do sample (greedy decoding)
False
Num beams
1
Max Soft Prompt Length n
20
Max Episode Length H
150
Mini-Batch Size
64
PPO Learning Rate
0.0002
Appendix
TABLE IX : Training parameters of our framework for the COCO dataset
Although Large Vision-Language Models (LVLMs) have demonstrated remarkable performance on downstream tasks, they frequently produce contents that deviate from visual information, leading to object hallucination. To tackle this, recent works mostly depend on expensive manual annotations and training cost, or decoding strategies which significantly increase inference time. In this work, we observe that LVLMs' attention to visual information is significantly enhanced when answering caption queries compared to non-caption queries. Inspired by this phenomenon, we propose Caption-guided Visual Attention Steering (CAST), a training-free, plug-and-play hallucination mitigation method that leverages the attention activation pattern corresponding to caption queries to enhance LVLMs' visual perception capability. Specifically, we use probing techniques to identify attention heads that are highly sensitive to caption queries and estimate optimized steering directions for their outputs. This steering strengthens LVLM's fine-grained visual perception capabilities, thereby effectively mitigating object hallucination. CAST reduced object hallucination by an average of 6.03% across five widely used LVLMs and five benchmarks including both discriminative and generative tasks, demonstrating state-of-the-art performance while adding little inference cost and preserving other foundational capabilities.
Qiming Li, Zekai Ye, Xiaocheng Feng +9
1Harbin Institute of Technology · 2Peng Cheng Laboratory · 3Harbin Institute of Technology, Shenzhen
The generation of factually incorrect objects, commonly known as object hallucination, remains a persistent challenge in Large Vision-Language Models (LVLMs). Current approaches to address this issue - ranging from expensive data-driven fine-tuning and high-latency contrastive decoding to rigid attention head truncation - frequently compromise either computational efficiency or the continuity of the model's feature space. To overcome these limitations, we introduce a novel, training-free inference strategy that operates as a region-aware adaptive weighting mechanism to dynamically correct semantic drift without relying on abrupt heuristic truncations. By computing an outlier-resistant statistical midpoint across various attention heads, we establish a stable anchor for reliable visual representations. We then utilize the inter-head disagreement mapped across regions to dynamically determine intervention budgets, gently suppressing hallucination-inducing attention paths through a continuous penalty modulation. This recalibration process effectively rectifies visual-semantic misalignments while fully preserving generative fluency and language priors. Comprehensive evaluations on standard multimodal benchmarks, including CHAIR, POPE, and MME, reveal that our strategy substantially curtails both instance- and sentence-level hallucinations. The results demonstrate state-of-the-art performance against contemporary baselines, confirming our method's efficiency and algorithmic robustness. Our code will be public.
Yuanzhi Xu, Qian Gao, Jun Fan +4
Qilu University of Technology (Shandong Academy of Sciences) Jinan, China · China Telecom Digital Intelligence Technology Co, Ltd Jinan, China · Shenyang Aerospace University Shenyang, China +1
Object hallucination is a significant challenge that hinders the application of large vision-language models (LVLMs) in practice. We hypothesize that one possible origin of hallucination is the model's tendency to prioritize text generation over meaningful interaction with images. To explore this, we examine the generation process and categorize text tokens into three groups: image-positive, invariant, and negative, based on their visual dependence on input image tokens. Our analysis reveals that most generated tokens are minimally influenced by the image information. This suggests that during the model's training stage, more emphasis is placed on learning how to follow textual instructions, rather than extracting information from images. Based on this finding, we propose adjusting the training weights of different tokens depending on their visual dependence to control hallucination. Additionally, we remove a portion of the training data that potentially contains more hallucinations as a data filtering strategy. Both methods achieve a reduction in hallucination without compromising response length or introducing additional computational costs during inference. We validate our methods across three LVLM variants, demonstrating the effectiveness and general applicability.
Meng Shen, Minghao Wu, Deepu Rajan
Nanyang Technological University · Monash University